Micro welding spot defect identification method based on computer vision

Abstract

The invention discloses a micro welding spot defect identification method based on computer vision, which comprises the following steps: S1, acquiring micro welding spot image data and preprocessing to generate a standardized image data set; s2, constructing a Gabor filter parameter space, and setting an initial value and a value range; s3, parameters are optimized based on a simulated annealing algorithm, Gabor features are extracted, and fitness is evaluated; s4, performing multi-scale and multi-direction filtering by using the optimal Gabor parameter to generate a high-dimensional texture feature vector; and S5, performing defect identification based on the high-dimensional texture feature vector, outputting defect types and positions, and performing feedback adjustment. According to the method, the simulated annealing algorithm and the Gabor filter are fused, so that efficient feature extraction and intelligent defect identification of the micro welding spot image are realized, and the detection accuracy and the automation level are improved.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a method for identifying micro-weld joint defects based on computer vision. Background Technology

[0002] With the rapid development of the microelectronics manufacturing industry, micro solder joints, as key connection components in integrated circuits and electronic packaging, directly affect the reliability and lifespan of electronic products. Existing methods for detecting micro solder joint defects mainly rely on manual visual inspection, traditional image processing techniques, and filtering and edge detection algorithms based on fixed parameters. These methods have significant shortcomings in terms of detection efficiency and automation. Manual inspection is easily influenced by subjectivity, leading to missed detections and misjudgments, making it difficult to meet the demands of large-scale, high-precision inspection. Traditional image processing methods often rely on manually set feature parameters, resulting in limited adaptability and generalization capabilities. Faced with complex and varied micro solder joint defect types and different manufacturing environments, they struggle to achieve high accuracy and robustness in automatic identification.

[0003] In recent years, Gabor filters have been widely used in image feature extraction and texture analysis due to their excellent spatial and frequency locality characteristics. However, traditional Gabor filters typically use a fixed set of parameters, making it difficult to adapt to different defect characteristics and varied real-world scenarios. Meanwhile, existing optimization methods, such as simulated annealing algorithms, suffer from low search efficiency and a tendency to get trapped in local optima in filter parameter selection, failing to achieve globally optimal configuration of Gabor filter parameters. Current technologies cannot effectively achieve efficient, accurate, and adaptive identification of micro-weld joint defects.

[0004] Therefore, how to provide a computer vision-based method for identifying micro-weld joint defects is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a computer vision-based method for identifying micro-weld joint defects. This invention fully integrates computer vision image feature extraction and intelligent optimization techniques. By adaptively optimizing the multi-scale, multi-directional parameters of the Gabor filter, it achieves efficient feature extraction and accurate automatic identification of micro-weld joint defect regions. The entire process of feature extraction, parameter optimization, identification, and dynamic feedback adjustment is described in detail. This method possesses advantages such as high detection accuracy, strong adaptability, high degree of automation in the identification process, and good stability, significantly improving the intelligence level and industrial practical value of micro-weld joint defect detection.

[0006] A method for identifying micro-weld joint defects based on computer vision according to an embodiment of the present invention includes the following steps:

[0007] S1. Collect image data of the micro-weld joints to be detected, obtain the original image dataset, preprocess the original image dataset to obtain the standardized image dataset;

[0008] S2. Based on the standardized image dataset, construct the Gabor filter parameter space and set the initial values ​​and ranges of the parameters;

[0009] S3. Using a standardized image dataset as input, simulated annealing algorithm is used to perform global search optimization of Gabor filter parameters, randomly perturb the parameters and generate a new parameter set, use the new parameter set to extract Gabor features from the standardized image dataset and evaluate the fitness value, determine whether to accept the new Gabor filter parameter set according to the Metropolis criterion, and gradually reduce the temperature until convergence to the optimal Gabor filter parameter set.

[0010] S4. Apply the optimal Gabor filter parameter set to the standardized image dataset, perform multi-scale and multi-directional Gabor filtering, obtain the filtering response at each scale and in each direction, and summarize all the response results to generate a high-dimensional texture feature vector.

[0011] S5. High-dimensional texture feature vectors are used for automatic defect identification, outputting the defect type, location, and related detection information of micro-welds. Based on the identification results, feedback of identification results and adaptive adjustment of parameters are performed to achieve dynamic optimization and continuous stable operation of the detection process.

[0012] Optionally, the original image dataset specifically includes original digital images of the micro-weld joint area acquired by a high-resolution imaging device.

[0013] Optionally, the preprocessing of the original image dataset specifically includes denoising, enhancing, cropping, and normalizing the images.

[0014] Optionally, S2 specifically includes:

[0015] S21. Based on the standard parameters of the Gabor filter—center frequency, directional angle, scale, and window size—add an adaptive response range factor. Among them, the center frequency controls the filter's sensitivity to different spatial frequencies, the direction angle determines the filter's response direction, the scale adjusts the width of the Gaussian envelope, the window size determines the filter's spatial dimensions, and the response range is determined by an adaptive factor. Dynamically adjust the spatial range of the Gaussian envelope in different regions. The range of values ​​is ;

[0016] S22. Introduce a direction-selective sensitivity coefficient in the parameter space. The filtered response is automatically weighted according to the defect distribution in different directions. The range of values ​​is and will Combined with the direction angle, the filter's response intensity in various directions can be flexibly adjusted;

[0017] S23. Add local suppression adjustment parameters to the parameter space. The filter output normalizes and suppresses the response intensity in local spatial regions. The range of values ​​is ;

[0018] S24. Analyze the edge density and texture complexity of each region in the standardized image dataset, and dynamically set the adaptive factor of the response range based on the analysis results. The specific values ​​enable the filter to increase its spatial range of action in areas with complex textures or defects, and to decrease its spatial range of action in uniform areas.

[0019] S25. Based on the main direction distribution characteristics of the micro-weld joint defects to be detected, dynamically set the direction selection sensitivity coefficient. The specific values ​​enhance the filter's response sensitivity to the main defect direction and suppress interference to non-target directions;

[0020] S26. Based on the response of the Gabor filter to the standardized image dataset, reasonably set the local suppression adjustment parameters. The specific values ​​are used to normalize and suppress high response points in local areas, reducing the impact of noise and abnormal responses on the overall characteristics.

[0021] S27. Adaptive factors for center frequency, orientation angle, scale, window size, and response range. Direction selection sensitivity coefficient and local inhibition regulation parameters They are included together in the parameter space of the Gabor filter, serving as all the structural parameters of the Gabor filter.

[0022] Optionally, S3 specifically includes:

[0023] S31. Set the center frequency, direction angle, scale, window size, response range adaptive factor, direction selection sensitivity coefficient and local suppression adjustment parameter in the Gabor filter parameter space as variables to be optimized, and set the initial value and allowable value range for each parameter respectively.

[0024] S32. Based on the standardized image dataset, select the initial Gabor filter parameter set and set the initial temperature and temperature reduction ratio of the simulated annealing algorithm.

[0025] S33. When generating a new parameter set, the parameter sets that have been accepted as optimal in the last n times are counted. For each parameter to be optimized, the arithmetic mean of the last n optimal parameter values ​​is calculated as the historical mean, and the standard deviation of the last n optimal parameter values ​​is calculated as the fluctuation range, where n is a positive integer. When generating a new parameter set, for each parameter, a new value is randomly selected with a probability of 0.7 from the interval between the historical mean minus the standard deviation and the historical mean plus the standard deviation, and a new value is randomly selected with a probability of 0.3 from the global value range allowed by the parameter, thus completing the generation of the new parameter set.

[0026] S34. Using the newly generated parameter set, Gabor features are extracted from the standardized image dataset, and the extracted features are compared with known defect types. By comparing the recognition results of each image with the actual defect annotation results, the proportion of samples identified as defects that are actually defects is calculated as the accuracy. The proportion of all actual defect samples that are correctly identified is calculated as the recall. The fitness value of the parameter set is calculated based on the accuracy and recall.

[0027] S35. Compare the fitness value of the new parameter group with the fitness value of the current optimal parameter group. If the fitness value of the new parameter group is higher than that of the current optimal parameter group, accept the new parameter group directly. If the fitness value of the new parameter group is lower than that of the current optimal parameter group, determine the acceptance probability based on the current temperature and the fitness difference: when the current temperature is high and the fitness difference is small, the probability of accepting the new parameter group is 90%; when the current temperature is low and the fitness difference is large, the probability of accepting the new parameter group is 10%; in other cases, the probability of accepting the new parameter group changes continuously between 10% and 90% depending on the temperature and the fitness difference.

[0028] S36. During the simulated annealing optimization process, continuously monitor the fitness improvement of the most recent several iterations. If the fitness improvement is less than the preset threshold for several consecutive iterations, increase the current temperature to several times the original temperature and increase the parameter perturbation amplitude. Conversely, decrease the temperature according to the set ratio.

[0029] S37. Every set number of iterations, perform clustering statistical analysis on the parameter groups that have been visited. If the density of the current parameter group in the clustering region exceeds the preset threshold, trigger a large step perturbation operation. That is, when randomly selecting a new parameter value within the allowed value range of each parameter, the perturbation amplitude is not less than 50% of the width of the allowed range of the parameter. The new parameter value has a greater than 50% probability of being selected at any position within the global allowed range.

[0030] S38. When the temperature drops to the minimum set value or the maximum set number of iterations is reached during the simulated annealing optimization process, the parameter search is terminated, and the currently obtained optimal Gabor filter parameter set is determined as the final optimization result.

[0031] Optionally, S4 specifically includes:

[0032] S41. The obtained optimal Gabor filter parameter set is adopted, which includes center frequency, direction angle, scale, window size, response range adaptive factor, direction selection sensitivity coefficient and local suppression adjustment parameter;

[0033] S42. Based on the center frequency, scale and window size in the optimal parameter set, construct Gabor filters of multiple scales on the standardized image dataset. For each scale, the filters are used with different parameter values ​​for spatial domain filtering.

[0034] S43. Select the sensitivity coefficient based on the direction angle and direction in the optimal parameter set, set multiple directions at each scale, and use different direction parameters for image filtering to enhance the response capability to features in different directions.

[0035] S44. In the process of multi-scale and multi-directional filtering, the response range adaptive factor dynamically adjusts the spatial range of each filter in different image regions.

[0036] S45. In the filter response output stage, the filter response value of each region is normalized and suppressed by the local suppression adjustment parameter to reduce the impact of abnormally high response values ​​on the overall characteristics.

[0037] S46. Summarize the Gabor filter response results at all scales and orientations, concatenate or synthesize the response values ​​of each image pixel under all filters to form a high-dimensional texture feature vector, and output the high-dimensional texture feature vector.

[0038] Optionally, S5 specifically includes:

[0039] S51. Compare the high-dimensional texture feature vector obtained after Gabor filtering with the standard data corresponding to each micro-weld point image to identify possible defect areas in the image.

[0040] S52. Based on the comparison results, determine the specific defect type for each micro-weld point image, including cold weld, porosity, crack, and bridging, and mark the position coordinates of each defect in the image.

[0041] S53. Organize the defect type and defect location information of each micro-weld point image, and output the defect category, corresponding image number and the specific coordinates of the defect in the image;

[0042] S54. Summarize and statistically analyze all identified detection data, and analyze the overall identification effect, including the accuracy of identification, missed detections, and false alarms.

[0043] S55. Based on the results of statistical analysis, provide feedback on the shortcomings in the current detection process, including unreasonable parameter settings or poor feature extraction results, and mark the parameters and processing steps that need to be optimized.

[0044] S56. In response to the shortcomings identified, the Gabor filter parameters and related image processing procedures are automatically adjusted, and the adjusted parameters are applied to the next round of detection.

[0045] The beneficial effects of this invention are:

[0046] This invention improves the efficiency and accuracy of automated identification of micro-solder joint defects by deeply integrating the spatial frequency characteristics of Gabor filters with the global optimization capabilities of simulated annealing algorithms. Compared with traditional image processing methods that rely on manual judgment or fixed parameter settings, this invention introduces a parameter adaptive mechanism, enabling Gabor filters to automatically adjust the optimal parameter set for different types and distributions of micro-solder joint defects, greatly enhancing the robustness and applicability of defect feature extraction. The improved simulated annealing algorithm enhances the global search capability and convergence efficiency of parameter optimization, effectively avoiding the problem of parameter selection easily getting trapped in local optima, and achieving high-precision defect detection in complex and variable manufacturing scenarios. This invention not only reduces manual intervention and improves the automation and intelligence level of the detection process, but also ensures the stability and reliability of the detection results. It is suitable for high-density, batch microelectronics manufacturing processes, providing an efficient, intelligent, and scalable technical solution for the field of micro-solder joint defect detection. Attached Figure Description

[0047] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0048] Figure 1 This is a flowchart of a micro-weld joint defect identification method based on computer vision proposed in this invention;

[0049] Figure 2 This is a schematic diagram illustrating the process of optimizing Gabor filter parameters through simulated annealing in a computer vision-based micro-weld joint defect identification method proposed in this invention. Detailed Implementation

[0050] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0051] refer to Figure 1A method for identifying micro-weld joint defects based on computer vision includes the following steps:

[0052] S1. Collect image data of the micro-weld joints to be detected, obtain the original image dataset, preprocess the original image dataset to obtain the standardized image dataset;

[0053] S2. Based on the standardized image dataset, construct the Gabor filter parameter space and set the initial values ​​and ranges of the parameters;

[0054] S3. Using a standardized image dataset as input, simulated annealing algorithm is used to perform global search optimization of Gabor filter parameters, randomly perturb the parameters and generate a new parameter set, use the new parameter set to extract Gabor features from the standardized image dataset and evaluate the fitness value, determine whether to accept the new Gabor filter parameter set according to the Metropolis criterion, and gradually reduce the temperature until convergence to the optimal Gabor filter parameter set.

[0055] S4. Apply the optimal Gabor filter parameter set to the standardized image dataset, perform multi-scale and multi-directional Gabor filtering, obtain the filtering response at each scale and in each direction, and summarize all the response results to generate a high-dimensional texture feature vector.

[0056] S5. High-dimensional texture feature vectors are used for automatic defect identification, outputting the defect type, location, and related detection information of micro-welds. Based on the identification results, feedback of identification results and adaptive adjustment of parameters are performed to achieve dynamic optimization and continuous stable operation of the detection process.

[0057] In this embodiment, the original image dataset specifically includes original digital images of the micro-weld joint area acquired by a high-resolution imaging device.

[0058] In this embodiment, the preprocessing of the original image dataset specifically includes denoising, enhancing, cropping, and normalizing the images.

[0059] In this embodiment, S2 specifically includes:

[0060] S21. Based on the standard parameters of the Gabor filter—center frequency, directional angle, scale, and window size—add an adaptive response range factor. Among them, the center frequency controls the filter's sensitivity to different spatial frequencies, the direction angle determines the filter's response direction, the scale adjusts the width of the Gaussian envelope, the window size determines the filter's spatial dimensions, and the response range is determined by an adaptive factor. Dynamically adjust the spatial range of the Gaussian envelope in different regions. The range of values ​​is ;

[0061] S22. Introduce a direction-selective sensitivity coefficient in the parameter space. The filtered response is automatically weighted according to the defect distribution in different directions. The range of values ​​is and will Combined with the direction angle, the filter's response intensity in various directions can be flexibly adjusted;

[0062] S23. Add local suppression adjustment parameters to the parameter space. The filter output normalizes and suppresses the response intensity in local spatial regions. The range of values ​​is ;

[0063] S24. Analyze the edge density and texture complexity of each region in the standardized image dataset, and dynamically set the adaptive factor of the response range based on the analysis results. The specific values ​​enable the filter to increase its spatial range of action in areas with complex textures or defects, and to decrease its spatial range of action in uniform areas.

[0064] S25. Based on the main direction distribution characteristics of the micro-weld joint defects to be detected, dynamically set the direction selection sensitivity coefficient. The specific values ​​enhance the filter's response sensitivity to the main defect direction and suppress interference to non-target directions;

[0065] S26. Based on the response of the Gabor filter to the standardized image dataset, reasonably set the local suppression adjustment parameters. The specific values ​​are used to normalize and suppress high response points in local areas, reducing the impact of noise and abnormal responses on the overall characteristics.

[0066] S27. Adaptive factors for center frequency, orientation angle, scale, window size, and response range. Direction selection sensitivity coefficient and local inhibition regulation parameters They are included together in the parameter space of the Gabor filter, serving as all the structural parameters of the Gabor filter.

[0067] In this embodiment, S3 specifically includes:

[0068] S31. Set the center frequency, direction angle, scale, window size, response range adaptive factor, direction selection sensitivity coefficient and local suppression adjustment parameter in the Gabor filter parameter space as variables to be optimized, and set the initial value and allowable value range for each parameter respectively.

[0069] S32. Based on the standardized image dataset, select the initial Gabor filter parameter set and set the initial temperature and temperature reduction ratio of the simulated annealing algorithm.

[0070] S33. When generating a new parameter set, the parameter sets that have been accepted as optimal in the last n times are counted. For each parameter to be optimized, the arithmetic mean of the last n optimal parameter values ​​is calculated as the historical mean, and the standard deviation of the last n optimal parameter values ​​is calculated as the fluctuation range, where n is a positive integer. When generating a new parameter set, for each parameter, a new value is randomly selected with a probability of 0.7 from the interval between the historical mean minus the standard deviation and the historical mean plus the standard deviation, and a new value is randomly selected with a probability of 0.3 from the global value range allowed by the parameter, thus completing the generation of the new parameter set.

[0071] S34. Using the newly generated parameter set, Gabor features are extracted from the standardized image dataset, and the extracted features are compared with known defect types. By comparing the recognition results of each image with the actual defect annotation results, the proportion of samples identified as defects that are actually defects is calculated as the accuracy. The proportion of all actual defect samples that are correctly identified is calculated as the recall. The fitness value of the parameter set is calculated based on the accuracy and recall.

[0072] S35. Compare the fitness value of the new parameter group with the fitness value of the current optimal parameter group. If the fitness value of the new parameter group is higher than that of the current optimal parameter group, accept the new parameter group directly. If the fitness value of the new parameter group is lower than that of the current optimal parameter group, determine the acceptance probability based on the current temperature and the fitness difference: when the current temperature is high and the fitness difference is small, the probability of accepting the new parameter group is 90%; when the current temperature is low and the fitness difference is large, the probability of accepting the new parameter group is 10%; in other cases, the probability of accepting the new parameter group changes continuously between 10% and 90% depending on the temperature and the fitness difference.

[0073] S36. During the simulated annealing optimization process, continuously monitor the fitness improvement of the most recent several iterations. If the fitness improvement is less than the preset threshold for several consecutive iterations, increase the current temperature to several times the original temperature and increase the parameter perturbation amplitude. Conversely, decrease the temperature according to the set ratio.

[0074] S37. Every set number of iterations, perform clustering statistical analysis on the parameter groups that have been visited. If the density of the current parameter group in the clustering region exceeds the preset threshold, trigger a large step perturbation operation. That is, when randomly selecting a new parameter value within the allowed value range of each parameter, the perturbation amplitude is not less than 50% of the width of the allowed range of the parameter. The new parameter value has a greater than 50% probability of being selected at any position within the global allowed range.

[0075] S38. When the temperature drops to the minimum set value or the maximum set number of iterations is reached during the simulated annealing optimization process, the parameter search is terminated, and the currently obtained optimal Gabor filter parameter set is determined as the final optimization result.

[0076] In this embodiment, S4 specifically includes:

[0077] S41. The obtained optimal Gabor filter parameter set is adopted, which includes center frequency, direction angle, scale, window size, response range adaptive factor, direction selection sensitivity coefficient and local suppression adjustment parameter;

[0078] S42. Based on the center frequency, scale and window size in the optimal parameter set, construct Gabor filters of multiple scales on the standardized image dataset. For each scale, the filters are used with different parameter values ​​for spatial domain filtering.

[0079] S43. Select the sensitivity coefficient based on the direction angle and direction in the optimal parameter set, set multiple directions at each scale, and use different direction parameters for image filtering to enhance the response capability to features in different directions.

[0080] S44. In the process of multi-scale and multi-directional filtering, the response range adaptive factor dynamically adjusts the spatial range of each filter in different image regions.

[0081] S45. In the filter response output stage, the filter response value of each region is normalized and suppressed by the local suppression adjustment parameter to reduce the impact of abnormally high response values ​​on the overall characteristics.

[0082] S46. Summarize the Gabor filter response results at all scales and orientations, concatenate or synthesize the response values ​​of each image pixel under all filters to form a high-dimensional texture feature vector, and output the high-dimensional texture feature vector.

[0083] In this embodiment, S5 specifically includes:

[0084] S51. Compare the high-dimensional texture feature vector obtained after Gabor filtering with the standard data corresponding to each micro-weld point image to identify possible defect areas in the image.

[0085] S52. Based on the comparison results, determine the specific defect type for each micro-weld point image, including cold weld, porosity, crack, and bridging, and mark the position coordinates of each defect in the image.

[0086] S53. Organize the defect type and defect location information of each micro-weld point image, and output the defect category, corresponding image number and the specific coordinates of the defect in the image;

[0087] S54. Summarize and statistically analyze all identified detection data, and analyze the overall identification effect, including the accuracy of identification, missed detections, and false alarms.

[0088] S55. Based on the results of statistical analysis, provide feedback on the shortcomings in the current detection process, including unreasonable parameter settings or poor feature extraction results, and mark the parameters and processing steps that need to be optimized.

[0089] S56. In response to the shortcomings identified, the Gabor filter parameters and related image processing procedures are automatically adjusted, and the adjusted parameters are applied to the next round of detection.

[0090] Example 1:

[0091] To verify the feasibility of this invention in practice, it was applied to the mobile phone motherboard production workshop of an electronics manufacturing company. The quality of micro-solder joints directly determines the overall reliability and market reputation of the product. Previously, this workshop mainly relied on manual visual inspection and traditional image processing methods for micro-solder joint defect detection. Operators had to take turns visually inspecting motherboards daily, with an average inspection time of 7 seconds per motherboard. Furthermore, subjective judgment errors were significant, and the varying inspection standards and experience levels among different employees led to frequent missed and false detections. Especially when facing small and complex defects such as porosity, cold solder joints, cracks, and bridging, the accuracy rate of manual inspection had consistently been below 87%, severely impacting product yield and resulting in frequent customer complaints and rework. Therefore, the company urgently needed a new method that could efficiently, stably, and automatically identify micro-solder joint defects.

[0092] In May 2025, the workshop introduced a computer vision-based micro-solder joint defect identification method proposed in this invention. A high-resolution industrial camera is installed at the end of the production line, automatically acquiring high-definition images of 1200 motherboards per hour. The system automatically performs preprocessing on each motherboard image, including noise reduction, enhancement, and ROI extraction. Subsequently, the system invokes an adaptive parameter optimization process, using a simulated annealing algorithm to globally search for the optimal Gabor filter parameter set within the parameter space. The process dynamically adjusts the search method based on historical optimal parameter distribution, temperature changes, and cluster distribution. After optimization, the system performs multi-scale, multi-directional Gabor filter feature extraction on each motherboard image, then compares it with standard defect samples, automatically outputting the type, location, and area of ​​each defect. All identification results are saved in real time for easy production statistics and traceability.

[0093] After three weeks of implementation, the proposed solution demonstrated significant advantages compared to traditional manual inspection and conventional fixed-parameter methods. The system takes an average of only 2.3 seconds to inspect each motherboard, achieving an overall accuracy rate of 98.5% and a recall rate of 97.9% for batch inspection of various defects such as cold solder joints, porosity, cracks, and bridging. Compared to manual inspection, the false negative rate decreased to 0.7%, and the false positive rate dropped to 1.2%. As the batches of raw materials and lighting conditions on the production line change, the system can adaptively optimize parameters daily, maintaining high-precision output. The product return rate decreased from 2.8% to 0.5%. Customer complaints decreased from an average of 17 per month to less than 2.

[0094] Table 1 Comparison of Micro-solder joint defect detection schemes

[0095] Testing Plan Average time (seconds / piece) Detection capacity (pieces / hour) Detection accuracy Detection recall False negative rate False alarm rate Return rate Customer complaints / month Manual inspection 7.0 480 86.7% 84.5% 5.8% 3.3% 2.8% 17 Fixed parameter algorithm 3.6 980 93.2% 91.8% 2.3% 2.0% 1.7% 7 Automatic identification scheme of the present invention 2.3 1200 98.5% 97.9% 0.7% 1.2% 0.5% 1～2

[0096] The data in Table 1 clearly demonstrates the significant advantages of the automatic identification scheme of this invention in the detection of micro-solder joint defects. In terms of detection efficiency, manual inspection of each motherboard takes an average of 7.0 seconds, with a detection capacity of only 480 pieces per hour; the fixed-parameter algorithm reduces the average time to 3.6 seconds, increasing the capacity to 980 pieces per hour; while the automatic identification scheme of this invention takes an average of only 2.3 seconds, with an hourly detection capacity of up to 1200 pieces, significantly improving production speed and alleviating the bottleneck of manual labor.

[0097] From the perspective of detection accuracy and robustness, manual detection achieves an accuracy of 86.7% and a recall of 84.5%, with a false negative rate as high as 5.8% and a false positive rate of 3.3%. Fixed-parameter algorithms show improved accuracy and recall, at 93.2% and 91.8% respectively, but still suffer from a false negative rate of 2.3% and a false positive rate of 2.0%. In contrast, the automatic identification scheme of this invention achieves an accuracy of 98.5%, a recall rate of 97.9%, a false negative rate reduced to 0.7%, and a false positive rate of only 1.2%, outperforming the previous two methods in all key detection metrics.

[0098] The changes in return rate and the number of customer complaints fully demonstrate the value of the method of this invention in engineering applications. Under the manual inspection scheme, the return rate is as high as 2.8%, with as many as 17 customer complaints per month. The fixed parameter algorithm reduces these figures to 1.7% and 7 complaints, respectively. In contrast, the automatic identification scheme of this invention has a return rate of only 0.5% and reduces the number of customer complaints to 1-2. This not only improves the company's delivery capabilities and brand reputation but also provides a solid quality assurance for large-scale automated production.

[0099] The automatic identification method of this invention is significantly superior to manual and traditional algorithm methods in terms of detection efficiency, accuracy, reliability and customer feedback. It can efficiently and intelligently support the batch and high-quality detection of micro solder joint defects, which is of great practical significance for promoting the automation upgrade of the electronics manufacturing industry.

[0100] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for identifying micro-weld joint defects based on computer vision, characterized in that, Includes the following steps: S1. Collect image data of the micro-weld joints to be detected, obtain the original image dataset, preprocess the original image dataset to obtain the standardized image dataset; S2. Based on the standardized image dataset, construct the Gabor filter parameter space and set the initial values ​​and ranges of the parameters; S3. Using a standardized image dataset as input, simulated annealing algorithm is used to perform global search optimization of Gabor filter parameters, randomly perturb the parameters and generate a new parameter set, use the new parameter set to extract Gabor features from the standardized image dataset and evaluate the fitness value, determine whether to accept the new Gabor filter parameter set according to the Metropolis criterion, and gradually reduce the temperature until convergence to the optimal Gabor filter parameter set. S4. Apply the optimal Gabor filter parameter set to the standardized image dataset, perform multi-scale and multi-directional Gabor filtering, obtain the filtering response at each scale and in each direction, and summarize all the response results to generate a high-dimensional texture feature vector. S5. High-dimensional texture feature vectors are used for automatic defect identification, outputting the defect type, location, and related detection information of micro-welds. Based on the identification results, feedback of identification results and adaptive adjustment of parameters are performed to achieve dynamic optimization and continuous stable operation of the detection process.

2. The micro-weld joint defect identification method based on computer vision according to claim 1, characterized in that, The original image dataset specifically includes original digital images of the micro-solder joint area acquired by a high-resolution imaging device.

3. The micro-weld joint defect identification method based on computer vision according to claim 1, characterized in that, The preprocessing of the original image dataset specifically includes denoising, enhancement, cropping, and normalization of the images.

4. The method for identifying micro-weld joint defects based on computer vision according to claim 1, characterized in that, S2 specifically includes: S21. Based on the standard parameters of the Gabor filter—center frequency, directional angle, scale, and window size—add an adaptive response range factor. Among them, the center frequency controls the filter's sensitivity to different spatial frequencies, the direction angle determines the filter's response direction, the scale adjusts the width of the Gaussian envelope, the window size determines the filter's spatial dimensions, and the response range is determined by an adaptive factor. Dynamically adjust the spatial range of the Gaussian envelope in different regions. The range of values ​​is ; S22. Introduce a direction-selective sensitivity coefficient in the parameter space. The filtered response is automatically weighted according to the defect distribution in different directions. The range of values ​​is and will Combined with the direction angle, the filter's response intensity in various directions can be flexibly adjusted; S23. Add local suppression adjustment parameters to the parameter space. The filter output normalizes and suppresses the response intensity in local spatial regions. The range of values ​​is ; S24. Analyze the edge density and texture complexity of each region in the standardized image dataset, and dynamically set the adaptive factor of the response range based on the analysis results. The specific values ​​enable the filter to increase its spatial range of action in areas with complex textures or defects, and to decrease its spatial range of action in uniform areas. S25. Based on the main direction distribution characteristics of the micro-weld joint defects to be detected, dynamically set the direction selection sensitivity coefficient. The specific values ​​enhance the filter's response sensitivity to the main defect direction and suppress interference to non-target directions; S26. Based on the response of the Gabor filter to the standardized image dataset, reasonably set the local suppression adjustment parameters. The specific values ​​are used to normalize and suppress high response points in local areas, reducing the impact of noise and abnormal responses on the overall characteristics. S27. Adaptive factors for center frequency, orientation angle, scale, window size, and response range. Direction selection sensitivity coefficient and local inhibition regulation parameters They are included together in the parameter space of the Gabor filter, serving as all the structural parameters of the Gabor filter.

5. The method for identifying micro-weld joint defects based on computer vision according to claim 1, characterized in that, S3 specifically includes: S31. Set the center frequency, direction angle, scale, window size, response range adaptive factor, direction selection sensitivity coefficient and local suppression adjustment parameter in the Gabor filter parameter space as variables to be optimized, and set the initial value and allowable value range for each parameter respectively. S32. Based on the standardized image dataset, select the initial Gabor filter parameter set and set the initial temperature and temperature reduction ratio of the simulated annealing algorithm. S33. When generating a new parameter set, the parameter sets that have been accepted as optimal in the last n times are counted. For each parameter to be optimized, the arithmetic mean of the last n optimal parameter values ​​is calculated as the historical mean, and the standard deviation of the last n optimal parameter values ​​is calculated as the fluctuation range, where n is a positive integer. When generating a new parameter set, for each parameter, a new value is randomly selected with a probability of 0.7 from the interval between the historical mean minus the standard deviation and the historical mean plus the standard deviation, and a new value is randomly selected with a probability of 0.3 from the global value range allowed by the parameter, thus completing the generation of the new parameter set. S34. Using the newly generated parameter set, Gabor features are extracted from the standardized image dataset, and the extracted features are compared with known defect types. By comparing the recognition results of each image with the actual defect annotation results, the proportion of samples identified as defects that are actually defects is calculated as the accuracy. The proportion of all actual defect samples that are correctly identified is calculated as the recall. The fitness value of the parameter set is calculated based on the accuracy and recall. S35. Compare the fitness value of the new parameter group with the fitness value of the current optimal parameter group. If the fitness value of the new parameter group is higher than that of the current optimal parameter group, accept the new parameter group directly. If the fitness value of the new parameter group is lower than that of the current optimal parameter group, determine the acceptance probability based on the current temperature and the fitness difference: when the current temperature is high and the fitness difference is small, the probability of accepting the new parameter group is 90%; when the current temperature is low and the fitness difference is large, the probability of accepting the new parameter group is 10%; in other cases, the probability of accepting the new parameter group changes continuously between 10% and 90% depending on the temperature and the fitness difference. S36. During the simulated annealing optimization process, continuously monitor the fitness improvement of the most recent several iterations. If the fitness improvement is less than the preset threshold for several consecutive iterations, increase the current temperature to several times the original temperature and increase the parameter perturbation amplitude. Conversely, decrease the temperature according to the set ratio. S37. Every set number of iterations, perform clustering statistical analysis on the parameter groups that have been visited. If the density of the current parameter group in the clustering region exceeds the preset threshold, trigger a large step perturbation operation. That is, when randomly selecting a new parameter value within the allowed value range of each parameter, the perturbation amplitude is not less than 50% of the width of the allowed range of the parameter. The new parameter value has a greater than 50% probability of being selected at any position within the global allowed range. S38. When the temperature drops to the minimum set value or the maximum set number of iterations is reached during the simulated annealing optimization process, the parameter search is terminated, and the currently obtained optimal Gabor filter parameter set is determined as the final optimization result.

6. The method for identifying micro-weld joint defects based on computer vision according to claim 1, characterized in that, S4 specifically includes: S41. The obtained optimal Gabor filter parameter set is adopted, which includes center frequency, direction angle, scale, window size, response range adaptive factor, direction selection sensitivity coefficient and local suppression adjustment parameter; S42. Based on the center frequency, scale and window size in the optimal parameter set, construct Gabor filters of multiple scales on the standardized image dataset. For each scale, the filters are used with different parameter values ​​for spatial domain filtering. S43. Select the sensitivity coefficient based on the direction angle and direction in the optimal parameter set, set multiple directions at each scale, and use different direction parameters for image filtering to enhance the response capability to features in different directions. S44. In the process of multi-scale and multi-directional filtering, the response range adaptive factor dynamically adjusts the spatial range of each filter in different image regions. S45. In the filter response output stage, the filter response value of each region is normalized and suppressed by the local suppression adjustment parameter to reduce the impact of abnormally high response values ​​on the overall characteristics. S46. Summarize the Gabor filter response results at all scales and orientations, concatenate or synthesize the response values ​​of each image pixel under all filters to form a high-dimensional texture feature vector, and output the high-dimensional texture feature vector.

7. The method for identifying micro-weld joint defects based on computer vision according to claim 1, characterized in that, S5 specifically includes: S51. Compare the high-dimensional texture feature vector obtained after Gabor filtering with the standard data corresponding to each micro-weld point image to identify possible defect areas in the image. S52. Based on the comparison results, determine the specific defect type for each micro-weld point image, including cold weld, porosity, crack, and bridging, and mark the position coordinates of each defect in the image. S53. Organize the defect type and defect location information of each micro-weld point image, and output the defect category, corresponding image number and the specific coordinates of the defect in the image; S54. Summarize and statistically analyze all identified detection data, and analyze the overall identification effect, including the accuracy of identification, missed detections, and false alarms. S55. Based on the results of statistical analysis, provide feedback on the shortcomings in the current detection process, including unreasonable parameter settings or poor feature extraction results, and mark the parameters and processing steps that need to be optimized. S56. In response to the shortcomings identified, the Gabor filter parameters and related image processing procedures are automatically adjusted, and the adjusted parameters are applied to the next round of detection.

Images